When the Cart and the Wallet Belong to the Platform: What Agent-Assisted Checkout Means for Retailers
For a generation, a retailer's website has been the place where the customer relationship is formed. The shopper arrives, signs in, builds a basket and pays, and the retailer keeps the identity, the payment data and the terms of the sale. Agent-assisted shopping puts a new party in the middle of that sequence. When the cart and the wallet belong to a platform rather than to the retailer, the first question stops being whether an agent can buy from you. It becomes who is still the customer's counterparty once it does. Our position is that retailers should answer that question deliberately, in writing, before a platform answers it for them.
What has actually been reported, and what has not
Start with the facts, because the topic attracts more speculation than evidence.
Google has launched Universal Cart in the US, with checkout through the Universal Commerce Protocol (UCP) and Google Pay, according to Google's own shopping announcement. Our reading of that launch, and the reading in the market intelligence we rely on, is that the wallet and the cart are becoming the control point for agent-assisted shopping, and that Google is backing its own protocol for checkout rather than waiting for a common one.
Three limits matter for a UK retailer.
- It is a US launch. UK availability has not been stated. Anything we say about UK retailers is argued from the shape of the move, not evidenced by a UK rollout.
- We have not seen the merchant terms. What a retailer receives about the customer, how orders are attributed and how returns and disputes are handled are all questions the launch coverage we have reviewed does not answer. We will not guess.
- One launch is one platform. Other platforms and networks are building their own routes, which we come to below.
There are also signals on the consumer side, and they point in two directions. Meta's Muse assistant passed 10 million downloads in the US and Canada in 35 days, according to Sensor Tower figures reported by PYMNTS. That shows how quickly a platform with an existing user base can distribute an assistant. It does not show how many people use it to buy: downloads are not active use, and the data covers North America only.
Consumer surveys reported in October found that willingness to use AI to research a purchase outpaces willingness to let it complete the purchase. That is a single set of survey reports, and we treat it as direction rather than measurement. Separately, Stripe's president has said that agentic-commerce activity "went vertical" from September, without disclosing figures. It is an executive's characterisation and we cannot size it.
Taken together, the plain summary is this: the infrastructure for delegated checkout is arriving faster than the evidence that shoppers will use it. That gap is the window a retailer has to prepare in.
Why the cart and the wallet are the control point
The rest of this article is analysis, and we label it as such.
In a conventional storefront journey, the retailer owns several things at once without ever having to name them. It owns the session, so it sees what the shopper looked at and abandoned. It owns the login, so it knows who the customer is. It owns the checkout page, so it chooses the payment methods, the delivery options and the upsell. It owns the confirmation, so it is the party that writes to the customer afterwards.
Agent-assisted shopping can pull those apart. If an assistant assembles the basket from several merchants, the session sits with the assistant. If the customer's credentials are held in a platform wallet, the login and the payment choice sit with the platform. If the platform's protocol defines how checkout is completed, the retailer's checkout page is no longer the place where terms are presented. Each of these is a possibility rather than a documented outcome of any one launch, but together they describe why we call the cart and the wallet the control point. Whoever holds them sets the terms on which everyone else participates.
This is the same dynamic that has played out in other channels, with one difference. The platforms are not converging on a shared standard. The industry intelligence we use describes UCP, Mastercard's Agent Pay for Machines, American Express's Intent Intelligence framework and the x401 and x402 approaches as coexisting without any one displacing the others, and it reads the picture as fragmentation widening rather than narrowing. For a retailer, that means there is no single integration to wait for, and a real risk of reworking an integration if an early commitment turns out to be to the scheme that does not win.
It also means that the retailer's leverage depends on being able to work with several routes without being captured by one. We return to that in the section on what to do now, and it is where agentic commerce architecture choices start to matter more than any single protocol.
Four things a retailer should insist on keeping
For each item, we set out the question to put to a platform and the position the retailer should take internally. None of this reflects a documented platform policy. It is a checklist for the conversation.
Identity: who is the customer?
The question for any platform is whether the retailer knows who bought. A retailer that receives an order from an assistant without a durable customer identifier cannot recognise a repeat customer, honour a loyalty entitlement or contact someone about a recall. The internal position should be that any agent-assisted channel must return enough identity, with the customer's consent, for the retailer to treat the order as a relationship and not a one-off transaction.
This is also a design question on the retailer's own side. If a customer can be recognised on your site only by a login that an assistant cannot present, you will need a way to link an agent-initiated order to an existing account. Work that out before the volume arrives, not after.
Payment data: what comes back after the transaction?
The question here is what the retailer receives when the wallet, rather than the retailer's own payment page, completes the payment. The relevant details include whether the retailer sees the payment method type, how refunds are routed, how a chargeback is evidenced and what happens to any tokens or credentials.
The retailer's internal position should be that it needs enough payment data to reconcile, to refund, to dispute and to detect fraud, and that it will not accept a channel where it carries the liability for a transaction it cannot evidence. We do not know the terms platforms will offer, so the practical step is to define your minimum before you are asked.
Terms: whose rules apply to the sale?
When an agent compares products across merchants, the terms of sale become part of what is being compared. Price, promotions, substitutions, delivery promises and returns policy may be read by the agent rather than by the shopper. If the platform's protocol defines how those are expressed, the platform is defining the vocabulary in which you compete.
The internal position should be that the retailer's terms are presented as the retailer's own, that substitutions and price changes require explicit rules, and that the retailer can see how its offer was represented. Without that, you cannot tell whether a lost sale was a pricing problem, a data problem or an agent that misread your returns policy.
Authority and liability: who said yes?
The fourth item is the one the wider market is already arguing about. Market intelligence across the last three reporting cycles describes the debate moving from whether agents can pay to who authorises a transaction and who is liable when it goes wrong. JPMorgan Chase has reportedly blocked outside AI agents from direct access to customer accounts and backed open-banking routes instead, as reported by Finextra citing CNBC. We read that as a signal that a large incumbent can insist on sanctioned, authorised access on its own terms.
The lesson for a retailer is modest but useful. Terms of access are negotiable by the party with leverage, and leverage comes from being clear about what you require. Agree in writing who proved the customer's authority, who bears the cost of an unauthorised order and how an agent's instruction is evidenced. American Express's Intent Intelligence framework, which links authenticated purchase intent to agent transactions, is one example of the industry working on that evidence. We have not assessed it in detail.
What a UK retailer should do now, and what to wait for
The sensible response to a US launch with unstated UK availability is neither to ignore it nor to rebuild the checkout. It is to separate work that pays off whichever platform wins from commitments that should wait.
Do now
- Map where you rely on owning the session. List the places in your business that assume you see the whole journey: personalisation, loyalty, fraud scoring, abandoned-basket recovery, customer service context. These are what you lose, or must replace, if the basket is assembled elsewhere.
- Write your minimum terms for any agent channel. Identity returned, payment data returned, terms presented as yours, liability allocated. Make this a one-page policy that commercial and legal teams have agreed, so that every platform conversation starts from the same position.
- Keep payment choice flexible. If your checkout is tightly coupled to one payment route, an agent-led journey will make that coupling expensive. A thin abstraction over digital wallets and payment methods lets you accept whichever credential an agent presents without reworking the core.
- Make your catalogue and policies machine-readable. Consistent, structured product data, delivery promises and returns rules serve every route, including your own site.
- Start with discovery, not delegated payment. The consumer survey evidence says shoppers are more comfortable with AI helping them research than with AI paying. Being accurately represented in research journeys is a lower-risk first step than enabling agent-completed orders.
Wait for
- UK availability and merchant terms. Do not plan around a launch whose UK status has not been stated.
- Convergence, or at least interoperability. With several protocols in play and none displacing the others, avoid exclusive commitments. Build to a protocol-agnostic internal model and add routes as the evidence justifies.
- Regulatory clarity where finance is involved. The FCA's 19 October speech on AI in financial services concerned financial firms rather than retailers, and it signalled that existing frameworks, including Consumer Duty and senior-manager accountability, apply to AI. Retailers offering regulated financial products, or working with partners who do, should ask their compliance advisers how that applies. We draw no further conclusion from it.
What we do not know
This article rests on one launch, reported by one platform, in one market. The market intelligence behind it is a single source, and we have not seen merchant terms, UK plans or any measured shopper behaviour on Universal Cart. We have not quoted any search-volume figure as evidence of demand. Where we have described how control might shift, we have written it as analysis, and a retailer should test each point against the actual terms offered.
Where Innovify fits
Innovify is a technology partner with deep expertise in banking, financial services, fintech and insurance. Our core nuance is that we bring fintech expertise to businesses that need to adopt fintech, including retailers and platforms that do not think of themselves as financial firms. The decisions in this article sit at the join between commerce and payments: what a checkout returns, how a wallet credential is handled and where authority is proven.
We design compliance in from day one, so identity, consent and audit points are part of the build rather than an addition. As an implementation partner across multiple fintech platforms, we help brands find a fast, compliant route to launching financial products. Our Agentic Commerce & Payments work covers the architecture questions above. If you would rather test an agent-assisted journey before committing, Innovify AI Labs lets enterprise teams run several proofs of concept in parallel, test with real users and finish with a business case ready for funding.
Frequently asked questions
Who controls checkout when an AI agent shops for a customer?
It depends on whose cart and wallet the agent uses, and that is the point at issue. If the platform holds the cart and the wallet credentials, the platform controls how checkout is completed unless the retailer's terms say otherwise. Retailers keep control by deciding in advance what identity, payment data and terms any agent-assisted channel must provide.
What is Google Universal Cart?
Google has launched Universal Cart in the US, with checkout through the Universal Commerce Protocol and Google Pay, according to Google's announcement. UK availability has not been stated. We have not seen the merchant terms, so retailers should avoid assuming how data, attribution or disputes are handled.
Will AI shopping agents take over the customer relationship from retailers?
Not automatically. Agents and platform wallets can take over parts of the relationship, such as the basket, the login and the payment choice, if retailers allow it. Survey reports suggest shoppers currently trust AI more for research than for completing purchases, so there is time to set terms before volume arrives.
Should a UK retailer integrate with the Universal Commerce Protocol now?
There is no evidence yet that an immediate integration is needed for UK retailers. UK availability has not been stated, and several competing protocols are emerging without convergence. A better first step is to build a protocol-agnostic internal model and decide your minimum terms, then add routes as they become relevant.
What should a retailer ask an agent platform before accepting orders through it?
Ask what customer identity the retailer receives, what payment data and refund routing it gets, how its terms and prices are represented to the agent, and who is liable for an unauthorised or misread order. Get the answers in writing. If a platform cannot answer them, treat the channel as an experiment rather than a dependency.
Conclusion
Platform-owned carts and wallets are being introduced, in the US at least, as the route to agent-assisted checkout, and each platform is backing its own protocol. The evidence that shoppers will delegate payment at scale is thin, which gives retailers time and a reason to be deliberate. Decide what you will not hand over: identity, payment data and your own terms, with authority and liability agreed in writing. Keep your integrations flexible, start with discovery rather than delegated payment, and test every assumption about control against the terms you are actually offered.
